Generation of Synthetic Ampacity and Electricity Pool Prices using Generative Adversarial Networks

This work explores the generation of synthetic time-series of dynamic thermal line rating data and Alberta Electric System Operator's hourly pool price data using Wasserstein Generative Adversarial Networks, as part of a larger study on transmission line reliability. The generation of synthetic data is required due to a limited size of the available dataset. Synthetic data can aid in training deep learning and reinforcement learning models. The data is generated for 100 time-steps and is evaluated using quantitative metrics and qualitative assessment methods. Results show that the maximum mean discrepancy loss stabilizes and the trained Wasserstein generative adversarial network is able to reproduce the desired frequency distributions as well as produce a good overlap in the principal component analysis decomposition between the real and synthetic data. The final inspection of the produced synthetic data on both datasets is satisfactory.

I. Introduction

Availability of data for training deep learning and reinforcement learning models can be a challenge in cases where data privacy and scale of data do not allow large volumes of data to be stored [1]. Additionally, some datasets rely on experts to design, which limits the amount of data in the datasets [2]. For simulations and scenario planning involving dynamic thermal line rating (DTLR), the privacy and scale, and for simulations with pool price data. the scale, may present an issue.